高级检索

外周血白细胞分类模型的建立与应用

Establishment and application of peripheral blood leukocyte classification model

    摘要
  • 摘要:
    目的 基于Swin Transformer模型进行外周血白细胞自动分类, 并与经典卷积神经网络模型ResNet进行比较。
    方法 以经典的卷积神经网络模型ResNet和新的Swin Transformer模型为网络原型进行训练，使用Cella Vision DI60自动分析仪采集白细胞图像，由2名经验丰富的检验人员确认细胞的类别标签。通过学习率衰减中的指数衰减方式使模型更快收敛，然后对2 788张白细胞图像进行测试。
    结果 ResNet模型对5类白细胞图像的平均测试准确率为95.2%, Swin Transformer模型则高达99.1%。Swin Transformer模型对中性粒细胞、嗜酸性粒细胞、嗜碱性粒细胞、淋巴细胞、单核细胞的识别准确率分别为99.8%、94.8%、97.5%、99.5%、93.8%。
    结论 Swin Transformer模型可减少计算量，更适用于白细胞分类识别，且准确率相较于ResNet模型更具优势。

     

    Abstract:
    Objective To automatically classify peripheral blood leukocytes based on Swin Transformer model, and to compare the difference between Swin Transformer model and ResNet that is a classical convolutional neural network model.
    Methods The classical convolutional neural network model ResNet and the new Swin Transformer model were used as network prototypes for training. White blood cell images were collected using the Cella Vision DI60 automatic analyzer, and the category labels of cells were confirmed by two experienced inspectors. The exponential attenuation, a learning rate attenuation method, was used to make the model converge faster. Then, 2, 788 leukocyte images were tested.
    Results The average test accuracy of ResNet for five kinds of leukocyte images was 95.2%, while that of Swin Transformer was as high as 99.1%. Among them, the recognition accuracy of Swin Transformer model is 99.8% for neutrophils, 94.8% for eosinophils, 97.5% for basophils, 99.5% for lymphocytes and 93.8% for monocytes.
    Conclusion Swin Transformer model reduces the amount of calculation, and is more suitable for leukocyte classification and recognition. Compared with ResNet, this model has more advantage in accuracy.

     

    • HTML全文
    • 参考文献(14)
    • 相关文章
    • 施引文献
  • 资源附件(0)

/

返回文章
返回